A member build from the labs, written up start to finish: the phone agent, the webhook, the two bugs that ate an evening, and what it cost to run.
Rebuilding a member's Live Lab project (with permission) – the smallest useful production agent we know how to teach: a phone voice agent that answers, takes a message, and emails the shop owner.
The shape of it
- A SIP number forwards to a voice platform that handles speech-to-text, an LLM for the conversation, and text-to-speech back.
- The agent has exactly one tool:
send_message(to, body)hitting a webhook. - The webhook writes the transcript to a database and emails the owner within about a minute of the call ending.
What broke (the useful part)
First, silence: the call connected, the greeting played, then nothing. The carrier was sending RTP media to the wrong address – a NAT problem that no amount of agent code fixes. The lesson: test the media path before the model path.
Second, double-emails: the webhook retried on a timeout that wasn't a failure. The lesson: every tool an agent can call must be idempotent, because agents will retry.
The honest cost
The member's usage that month: a handful of dollars in carrier minutes and pennies per call in inference – but we won't dress that up as a business plan. What it bought was understanding: where voice breaks, where webhooks break, and which failures no model tutorial prepares you for.
The build isn't the demo. The build is the two evenings after the demo stops working.